This repository contains the implementation of a vision-based pursuer drone that tracks an evader in 3D using monocular/depth based strategies. The code is part of Siddharth Anand's Final Thesis Project at IIT Bombay under the guidence of Prof. Debraj Chakraborty.
Devloped by Siddharth Anand with contributions from Anjaneya Damle.

• Simulate real-time drone pursuit in a 3D AirSim environment
• Depth estimate for chase using 2 sensing strategies
  • Strategy 1: monocular camera and drone width ratio
  • Strategy 2: using depth perspective airsim-API
• Maintain evader visibility using PID-based gimbal control
• Handle edge cases like gimbal lock with recovery logic

• ACGC_3d*.py – Main simulation scripts
• drone_boundingbox.py – Object localization utilities
• requirements.txt – Python packages used

• Python 3.8+
• Cosys AirSim + Unreal Engine (Windows)
• OpenCV, NumPy, matplotlib

• Ensure settings.json replaces the default file in your AirSim Folder
• Now with the custom settings you can use the AirSim plugin in any unreal project.
• Start the simulation in Unreal Engine and then run the code by:

python ACGC_3d-depth_cam.py --EVADER_MOTION --TRACKING_METRICS --STRATEGY_FOR_DEPTH


EVADER_MOTIONS : --v_straight (Moves along vertical line) 
                 --circle (Moves in circular path with constant upward velocity [Spiral])
                 --depth (Moves into the plane [Depth axis])


TRACKING_METRICS : --plot (Plots angle error of heading vector and distance between evader and persuer)
                   --3d_track (3D Mapping  of chase)
                   (Can use both together)

STRATEGY_FOR_DEPTH : --strategy 1 (Monocular Camera with metric scaling based approach)
                     --strategy 2 (Depth camera based approach, Uses Airsim depth perspective API) [DEFAULT]
Example:
python ACGC_3d-depth_cam.py --v_straight --3d_track --plot --strategy 2

A complete set of results based on our testing can be found here